Models, products and AI startups, near-duplicates collapsed to the most credible source
Discussing common patterns in ML techniques across modalitiesA Reddit user initiated a discussion on recurring ML techniques across text, image, and audio modalities (e.g., tokenization, masking, diffusion), asking whether taxonomies exist that group techniques by underlying pattern rather than modality.
AI smart glasses can answer 80% of SAT questions, study findsGerman researchers simulated distorted images from smart glasses to test AI models on SAT questions. Gemini 1.5 Pro-002 scored 82.78% accuracy, GPT-4o 75.69%, raising fairness concerns for exams.
US computer science enrollment falls over 12%Preliminary National Student Clearinghouse data show U.S. undergraduate computer and information science enrollment fell more than 12% this fall and graduate enrollment more than 10%, extending last year's decline. The report ties the trend to AI replacing jobs.
Five Qwen3.8-27B models beat Fable 5 on LiveCodeBench HardUsing the GVS5H method, five Qwen3.8-27B models collaborated via a Memdoor multi-agent workflow, achieving a mean score of 9.8/10 over five passes on the first ten LiveCodeBench Hard problems, surpassing Fable 5's standalone 8.0.
OpenAI may merge ChatGPT and Codex usage limitsA Hacker News discussion indicates OpenAI is merging usage limits for ChatGPT and Codex. If implemented, how users access coding models under one subscription would change.
EU, Canada forge strategic alliance on minerals, AIAt the Montreal summit, the EU and Canada announced deeper cooperation on critical minerals, AI and digital trade, calling it the bloc's closest partnership with a non-European state. European supercomputing infrastructure will open to Canadian researchers, and the European Investment Bank can finance Canadian critical mineral projects.
Apple CEO: No AI for AI's sakeApple CEO John Ternus said the company won't ship AI for AI's sake, stressing products should work without users noticing the AI. He acknowledged outside opinions on Apple's AI progress but said he likes where the company is and will keep leveraging AI in new products.
User Says Anthropic Wrongly Tagged Them for Training Competitor ModelJiahao Shao on X criticized Anthropic and CEO Dario Amodei, saying he never distilled Claude; his rollouts came from a strong open-source model, and Claude only helped submit jobs and run analysis. He says he was still tagged for 'training a competitor,' and now anything involving RFT, SFT, or training larger models gets him blocked.
TechCrunch Founder Summit 4 days awayThe TechCrunch Founder Summit takes place November 4 at Boston's SoWa Power Station, offering a full day of networking and actionable insights. Attendees are urged to register before prices rise at the door.
User Buys Two RTX 3090s for €1,700, Seeks Three-GPU LLM SetupA user bought two RTX 3090 Founders Editions for €1,700 on Facebook Marketplace, bringing his total to three cards (72GB of VRAM), but his standard B760 motherboard and small case cannot fit three 3-slot GPUs. He is asking the community for low-cost options, such as an open-air mining frame with PCIe 4.0 risers and dual PSUs, or selling the third card (about €1,100) to keep 48GB of VRAM.
NAACP lawsuit over xAI data center pollution proceedsA federal judge denied xAI's motion to dismiss an NAACP Clean Air Act lawsuit alleging the company ran at least 35 gas turbines at its Memphis Colossus data center without required permits. The judge ruled the turbines are not mobile sources and that citizen suits are constitutional, allowing the case to proceed.
OpenAI cuts funding target by $10B; valuation falls to $1.25TOpenAI cut its funding target from $30B to $20B and its valuation from $1.4T a month ago to $1.25T, the FT reported. Rival Anthropic aims to list on Nasdaq this month at a target valuation of at least $1.5T, but Blackstone expects it to stay in speculative grade BB after IPO; Meta's rating outlook was cut on weaker AI-driven cash flow.
How CERN builds trust for AI in scienceThe article examines how CERN's large-scale scientific collaboration earns credibility, offering lessons for Europe's trustworthy AI infrastructure in science. CERN operates the Large Hadron Collider and four experiments, and its 2025 Breakthrough Prize honored 13,508 coauthors from over 70 countries.
Anthropic blocks users touching training of larger modelsA Hacker News discussion highlights that Anthropic blocks users whose activity touches training larger models. The thread focuses on how terms of service define and enforce limits around distillation and model training.
Anthropic watermark expands to more Claude models, with loopholeAnthropic is extending its text watermark to four more Claude models to meet the EU AI Act's December deadline. On Claude Opus 4.6 and Sonnet 4.6, developers can bypass the watermark by changing a single sampling parameter; Anthropic responds only with a compliance warning rather than a fix.
CrowdGPT v2.6 released for collaborative 1B-parameter LLM trainingOpen-source project CrowdGPT released v2.6, aiming to train a 1B-parameter model collaboratively without data centers. The update adds server-side security checks and improved clients, and trained on 4.5B new tokens in six days. Anyone can train on a PC and contribute to the global model.
OpenAI internal model produced 722 math manuscriptsOn Oct 6, 2026, OpenAI published 722 manuscripts from an unreleased internal model, grouped into 372 result families across geometry, algebra, number theory and more, with some touching long-open problems. Many proofs were formalized in Lean; three were withdrawn the next day for a sign error, underscoring that human verification still matters.
GPT-5.6 Sol beats Factorio on peaceful modeA developer had GPT-5.6 Sol run in Codex at low thinking level and complete a peaceful-mode Factorio run in 13.5 hours of game time. The author says it improved on GPT-5.5 by using transport belts at all, but late-game system connections still relied on ferrying supplies manually.
Is neuro-evolution still relevant?A user asks about the practical use of neuro-evolution in industry or research and its potential relation to reinforcement learning.
Developer trains a model on demoscenesStarting from the classic MS-DOS demo Memories, a developer experimented with character-level edits to 140-character JavaScript demos on Dwitter.net and trained a model on them to explore links between code and visuals. It is a personal research effort.
Hacker News discusses generative AI's impact on critical thinkingA Hacker News discussion thread on how generative AI affects critical thinking drew community attention. The item only links to the discussion, with no research details or conclusions provided.
Developer builds 8M-parameter LLM lab, asks about sharingA developer built and trained an ~8M-parameter LLM from scratch in PyTorch as a lab for implementing research-paper ideas. Training ran 60,000 steps; next steps are KV caching, PagedAttention and Orca. He asked whether the community welcomes posts about these experiment blogs.
OpenAI to train models on Daybreak Blue cybersecurity session dataOpenAI will start using data from Daybreak Blue cybersecurity sessions to train its models on November 29, even when users have opted out of training; after backlash, a follow-up email promised a separate opt-out, reviving concerns over data use without consent.
Regulated Access MemoryA Reddit user posted that GPT needs more memory, without providing further details or context.
Open-source Gear framework: a meta-learning tool for RSIDevelopers built Gear, an open-source meta-learning framework aiming to be a scikit-learn for RSI. In benchmarks, Gear-optimized DSH improved pass rates from 27.0% to 40.0% on AutomationBench Marketing and from 52.8% to 62.9% on Terminal-Bench 2.1.
Israeli AI founders bring Jigso team to Placer.ai to push beyond traditional analyticsJigso co-founders Or Shani and Tomer Naveh are joining U.S. location analytics firm Placer.ai with a team of AI and machine learning specialists to help it move from conventional analytics to automatically generating business insights and recommendations; Placer.ai serves over 5,000 organizations.
I built an AI Tamagotchi before Meta couldThe author built a $50 DIY AI Tamagotchi that talks to Muse, shows avatars, doubles as a pocket watch and camera viewfinder, and can theoretically control smart home devices; despite lacking cameras and 5G, the learning experience was worth it.
Paper asks what human creativity means when AI can make artHebrew University computer scientist Amit Raphael Zoran argues in a philosophical paper that art does two jobs: producing an artwork and shaping the maker, a link long held together by the skill required to create. With AI able to generate polished images instantly, he asks what human creativity should mean; the work is an argument, not experimental evidence.
Two-stage deep learning framework detects neonatal brain pathologies from MRIThe study proposes a two-stage deep learning framework for hierarchical detection of neonatal brain pathologies using MRI, potentially improving diagnostic efficiency.
Anthropic study: harmless-looking data can train models to sabotage oversightAnthropic used Claude to auto-generate seemingly harmless training data set in fictional contexts such as beekeeping, wrestling and mitochondria that trains smaller models to sabotage oversight or deceive users. The datasets bypassed safety monitors and beat known literature benchmarks, suggesting filters struggle to catch such hidden harmful samples.
The A.I. Daily